This paper discusses structure model based correction of thermal induced errors at machine tools. Using a machine model evaluated in thermal real-time, the thermal induced errors at the tool center point (TCP) are calculated based on information gotten from the machine control (e.g., axes velocities, positions, and motor currents) and ambient temperature. The machine model describes the physical relationships and considers the structure and structural variability resulting in traverse movements of the feed axes – the so-called structure model. To create this, finite elements are used as thermal and thermo-elastic models, and model order reduction (MOR) techniques are used to enable the calculation of high-resolution models in thermal real-time. Subsequent parameter updates can improve the accuracy of the initial parameter set of thermal models. A systematic procedure developed for this purpose and its application to a demonstrator machine are presented. For the update, parameters are selected which can change over the operating time, e.g., due to wear. Temperature sensor positions are chosen, sensitive to changes in these parameters. Simulations with parameters varied in a plausible range are used to determine whether parameter optimizations are reasonable. The parameter optimization runs in a trusted execution environment (TEE) on a server in parallel to the calculation of the correction model on the machine control. The confidential input data of the model and the model itself have to be protected from unauthorized access. The efficient model calculation and parameter optimization in a secure server environment leads to an adaptive thermal model (digital twin).